The question at this stage Will anyone pay, and who are they?
Record now, or lose it: a dated history of each customer's plan and status changes. Once overwritten, it often cannot be rebuilt.
The five steps at this stage
- Step 1Choose the problem
Write the one question that leads to a first paying customer and the decision it informs.
If you skip it: You build and measure things that never lead to a first customer.
- Step 2Collect what you need
Give each customer one ID from the first sign-up or deal. Mark test accounts when you create them. Keep a dated history of plan and status changes, mark deleted records instead of removing them, record a few key product actions, and restore a backup once. Then stop.
If you skip it: Your first customers' history is gone, and it often cannot be rebuilt later.
- Step 3Choose the number
Not yet: a full list of numbers. For now, write the one or two numbers that answer your question in a one page glossary, each with its decision and why you chose it.
If you skip it: The first definition belongs to whoever builds the first chart.
- Step 4Count it right
Leave test and internal accounts out of every number from day one. Point the agent at a copy of the data, read-only, not the live database.
If you skip it: Your own test sign-ups look like early traction, and an agent query can slow the live product.
- Step 5Read it right
Not yet: rates. With so few customers a percentage is noise, so read each account and check every agent answer against what you know about your customers.
If you skip it: A rate on a handful of people reads like a trend.
What can wait
- You can skip new dashboards until the question and its decision are written down.
- Before your first paying customers you can skip a separate analytics database. One customer ID, dated history and a backup you have restored once come first.
- Until growth is repeatable you can skip new monitoring tools. A copy of the data, a test flag and one second source to compare against come first.
- Safe by default
Collect only what you need, mark personal details, keep keys out of the code, and give each person their own login.
- Who runs it
The founder runs it, with an AI agent.
An experienced data person reviews the setup once, for a few hours.